Self-adaptive routing method and system for large model agent

Abstract

The invention provides a large model agent-oriented adaptive routing method and system, and relates to the field of large model agents, and the method comprises the steps: obtaining the information of a to-be-processed task of a certain edge agent, and evaluating whether the capability of the edge agent for processing the to-be-processed task is greater than a prediction processing threshold value or not; the edge agent sends a resource request to an adjacent node, receives a resource state reported by the adjacent node and determines the selected adjacent node; the edge agent transmits original data of the task to the selected adjacent node; the selected adjacent node processes the received original data and extracts key information from a processing result, and the selected adjacent node transmits the key information to the central cloud node through a communication channel between the selected adjacent node and the central cloud node, so that timely transmission and processing of the key information are guaranteed, and the security of the central cloud node is improved. And the response timeliness and robustness of the distributed artificial intelligence system in a sudden high-priority task scene are improved.

Description

Technical Field

[0001] This application relates to the field of large model intelligent agent technology, and more specifically, to an adaptive routing method and system for large model intelligent agents. Background Technology

[0002] In distributed artificial intelligence systems, especially those responsible for managing critical infrastructure such as urban transportation, a large number of edge computing nodes and central cloud computing facilities are deployed. Various intelligent agents run on these nodes, working collaboratively to achieve real-time monitoring and scheduling of urban traffic. However, when edge agents handle sudden high-priority tasks, they may encounter situations where their computing resources are insufficient to meet the demands. Under the complex situation of dual bottlenecks in computing and communication, existing static routing strategies and simple failover mechanisms are often ineffective, leading to delays in critical information processing and affecting the timeliness of emergency response.

[0003] For example, in the intelligent traffic management system of a large city, when a serious multi-vehicle pileup occurs, the edge agent responsible for monitoring that section of road immediately identifies the sudden event. Unlike conventional traffic flow statistics tasks, the nature of this task has fundamentally changed. The agent needs to perform detailed analysis of the accident scene. This task requires computing resources far exceeding the upper limit of the agent's own hardware configuration, causing its local processing unit to instantly reach 100% occupancy, making it impossible to complete the analysis within the required timeframe.

[0004] According to the system's pre-set contingency plan, when the edge agent's computing power is insufficient, it should directly upload the original high-definition video stream to the central cloud agent, utilizing the cloud's powerful computing capabilities for analysis. However, an unexpected situation prevented this plan from being executed. The backbone fiber optic network experienced a surge in load due to the need to transmit a large amount of high-definition video. At the same time, congestion on the public network also affected the communication quality of the backup lines. Both uplink paths to the central cloud experienced severe transmission bottlenecks, the data packet loss rate increased sharply, and the latency became unacceptable. Summary of the Invention

[0005] This application discloses an adaptive routing method and system for large-scale intelligent agents, aiming to solve the technical problem that large-scale sudden traffic events cause traditional prediction mechanisms to fail, thereby affecting the energy supply of key transportation units and even potentially causing grid instability.

[0006] The technical solution of this application is as follows:

[0007] In a first aspect, this application discloses an adaptive routing method for large-scale intelligent agents, the method comprising:

[0008] Obtain information about the task to be processed by a certain edge agent, and evaluate whether the edge agent's ability to process the task is greater than the predicted processing threshold. At the same time, evaluate whether the transmission status of all communication links between the edge agent and the central cloud node is blocked.

[0009] When the edge agent's ability to process the task to be processed is greater than the predicted processing threshold, and the transmission status of all communication links is blocked, the edge agent sends a resource request to the neighboring nodes and receives the resource status reported by the neighboring nodes. Based on the resource status, it selects a neighboring node as the processing node for the task to be processed and determines the selected neighboring node.

[0010] The edge agent establishes a direct communication channel with the selected neighboring node and transmits the raw data of the task to the selected neighboring node;

[0011] The selected neighboring node processes the received raw data and extracts key information from the processing result. The amount of key information is less than the amount of raw data.

[0012] The selected neighboring node transmits this critical information to the central cloud node through the communication channel that exists between it and the central cloud node.

[0013] Secondly, this application also discloses an adaptive routing system for large-scale intelligent agents, the system comprising:

[0014] The information acquisition and evaluation module is used to acquire information about the task to be processed by a certain edge agent, and evaluate whether the edge agent's ability to process the task is greater than the predicted processing threshold. At the same time, it evaluates whether the transmission status of all communication links between the edge agent and the central cloud node is blocked.

[0015] The neighbor node selection module is used to send resource requests to neighbor nodes and receive resource status reports from neighbor nodes when the edge agent's ability to process the task to be processed is greater than the predicted processing threshold and the transmission status of all communication links is blocked. Based on the resource status, the edge agent selects a neighbor node as the processing node for the task to be processed and determines the selected neighbor node.

[0016] The direct communication channel establishment and data transmission module is used to establish a direct communication channel between the edge agent and the selected neighboring node, and to transmit the raw data of the task to the selected neighboring node;

[0017] The data processing and key information extraction module is used by the selected neighboring node to process the received raw data and extract key information from the processing result. The amount of key information is less than the amount of raw data.

[0018] The critical information transmission module is used by the selected neighboring node to transmit the critical information to the central cloud node through the communication channel that exists between the neighboring node and the central cloud node.

[0019] Beneficial effects

[0020] This application discloses an adaptive routing method for large-scale intelligent agents. By acquiring information about the tasks to be processed by the edge agent and evaluating its own processing capabilities and the status of its communication link with the central cloud node, when the edge agent's processing capabilities are insufficient and the communication link is blocked, it intelligently sends resource requests to neighboring nodes and selects a suitable neighboring node as the processing node. Subsequently, the edge agent establishes a direct communication channel with the selected neighboring node to transmit the original task data. The selected neighboring node processes the original data, extracts key information smaller than the original data, and transmits this key information through its communication channel with the central cloud node. This method effectively solves the problem in existing technologies where static routing strategies and simple failover mechanisms fail when edge agents face the dual bottlenecks of insufficient computing resources and blocked uplink communication links. By intelligently routing tasks to neighboring nodes for preliminary processing and extraction of key information, this application significantly reduces the demand for communication bandwidth, thereby ensuring the timely transmission and processing of key information and improving the responsiveness and robustness of distributed artificial intelligence systems in scenarios with sudden high-priority tasks. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of an adaptive routing method for large-scale intelligent agents provided in this application.

[0022] Figure 2 This is a schematic diagram of an adaptive routing system for large-scale intelligent agents provided in this application. Detailed Implementation

[0023] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0024] Reference Figure 1 The diagram illustrates an embodiment of an adaptive routing method for large-scale intelligent agents according to an embodiment of the present invention, which may specifically include the following steps:

[0025] S101, obtain information about the task to be processed by a certain edge agent, and evaluate whether the edge agent's ability to process the task to be processed is greater than the predicted processing threshold, and at the same time evaluate whether the transmission status of all communication links between the edge agent and the central cloud node is blocked.

[0026] S102, when the edge agent's ability to process the task to be processed is greater than the predicted processing threshold, and the transmission status of all communication links is blocked, the edge agent sends a resource request to the neighboring node, receives the resource status reported by the neighboring node, selects a neighboring node as the processing node for the task to be processed according to the resource status, and determines the selected neighboring node.

[0027] S103, the edge agent establishes a direct communication channel with the selected neighboring node and transmits the raw data of the task to be processed to the selected neighboring node;

[0028] S104, the selected neighboring node processes the received raw data and extracts key information from the processing result, wherein the amount of key information is less than the amount of raw data;

[0029] S105, the selected neighboring node transmits the key information to the central cloud node through the communication channel that exists between it and the central cloud node.

[0030] This application enables edge agents to intelligently route tasks to neighboring nodes for processing when their own processing capabilities are limited and communication links with the central cloud node are blocked. By extracting key information, the amount of data transmitted is reduced, thereby effectively addressing the dual bottlenecks of computing and communication and ensuring the timely processing of high-priority tasks.

[0031] To better understand the adaptive routing method proposed in this application, some key terms and implementation environments are first explained. In this application, "edge agent" refers to an intelligent entity deployed at the network edge, possessing certain computing, storage, and communication capabilities, such as a smart camera, sensor gateway, or small edge server. These entities can initially process local data and perform specific tasks. "Central cloud node" represents a centralized cloud computing platform with powerful computing and storage capabilities, responsible for global scheduling, complex analysis, and data storage. "Nearby nodes" refer to other edge agents or edge servers geographically or topologically close to the edge agents, enabling direct communication and collaboration. "Predicted processing threshold" is a preset value used as a benchmark to measure the capability required for an edge agent to handle a specific task. When the actual processing capability of an edge agent is lower than this threshold, it indicates that it may be unable to complete the task independently. "Blocked state" refers to a state where the communication link experiences a significant decrease in transmission rate or an increase in packet loss rate due to excessive traffic, faults, or other reasons, failing to meet normal communication needs. The implementation environment of this application is typically a distributed artificial intelligence system, such as an intelligent traffic management system or a smart city monitoring system, which includes multiple edge agents, nearby nodes, and one or more central cloud nodes.

[0032] This application provides an adaptive routing method for large-scale intelligent agents, the core of which lies in intelligently determining the task processing path to cope with resource constraints and network congestion in edge computing environments.

[0033] First, the method includes acquiring information about a task to be processed by an edge agent, assessing whether the edge agent's ability to process the task exceeds a predicted processing threshold, and evaluating whether the transmission status of all communication links between the edge agent and the central cloud node is blocked. In practical applications, the edge agent can continuously monitor its own resource usage, such as CPU utilization, memory usage, and GPU load, and, combined with the type and complexity of the task to be processed, predict its required processing capacity using pre-trained models or heuristic algorithms. For example, for a video analysis task, the required computing resources can be predicted based on the video resolution, frame rate, complexity of the analysis algorithm, and historical processing data. Simultaneously, the edge agent can periodically send probe packets to the central cloud node or use network management protocols (such as SNMP) to obtain the real-time transmission status of all communication links with the central cloud node, including but not limited to indicators such as bandwidth, latency, and packet loss rate, and determine whether the links are blocked based on these indicators. For example, when the packet loss rate exceeds a certain preset percentage or the latency exceeds a certain number of milliseconds, the link can be determined to be blocked.

[0034] Secondly, when the edge agent's ability to process the task exceeds the predicted processing threshold, and all communication links are blocked, the edge agent sends a resource request to neighboring nodes and receives resource status reports from these nodes. Based on the resource status, it selects a neighboring node as the processing node for the task. When the edge agent determines that it cannot effectively process the task and cannot communicate normally with the central cloud node, it actively seeks assistance from neighboring nodes. The edge agent can broadcast resource request messages to all or some of its pre-configured list of neighboring nodes. These messages may include information such as the type of task, the required computing resources, and task priority. Upon receiving a resource request, a neighboring node assesses its current resource status (e.g., number of idle CPU cores, available memory, network bandwidth, etc.) and task load, and reports the assessment results to the requesting edge agent. After receiving resource status reports from multiple neighboring nodes, the edge agent can employ various strategies to select the optimal neighboring node. For example, you can simply select the neighboring node that reports the most idle resources, or select the nearest neighboring node that meets your resource requirements, or comprehensively consider factors such as resource quantity, network latency, and task priority, and select the neighboring node with the highest comprehensive score as the selected neighboring node through a weighted scoring mechanism.

[0035] Next, the edge agent establishes a direct communication channel with the selected neighboring node and transmits the raw data of the task to be processed to the selected neighboring node. Once the selected neighboring node is determined, the edge agent establishes a point-to-point direct communication channel with it. This channel can be an encrypted connection based on the TCP / IP protocol, or it can utilize the existing low-latency, high-bandwidth local area network connection in the edge network. After the channel is established, the edge agent transmits the raw data of the task to be processed, such as high-definition video streams and raw sensor data, to the selected neighboring node through this direct communication channel. This direct communication avoids the additional latency and network congestion that may be caused by data detouring through the central cloud node.

[0036] Next, the selected neighboring nodes process the received raw data and extract key information from the processing results. The amount of key information is smaller than the amount of raw data. After receiving the raw data, the selected neighboring nodes utilize their own computing resources to process the data. For example, if the raw data is a video stream, the selected neighboring nodes can run video analysis algorithms to identify specific events in the video (such as accidents or abnormal behavior) or extract structured metadata (such as vehicle type, license plate number, timestamp, location information, etc.). After processing, the selected neighboring nodes do not send back all processing results or raw data, but instead extract the "key information" that is most valuable to the central cloud node. This key information is highly abstracted and compressed data, and its amount is much smaller than the amount of raw data. For example, for a traffic accident, the key information may only be the time and location of the accident, the number of vehicles involved, the preliminary damage assessment, and the judgment of whether anyone was injured, rather than a high-definition video of the entire accident process.

[0037] Finally, the selected neighboring nodes transmit the critical information to the central cloud node through the communication channel they share with it. After extracting the critical information, the selected neighboring nodes utilize their existing communication channel with the central cloud node to transmit this data. Because the amount of critical information is significantly reduced, even though the communication channel between the selected neighboring nodes and the central cloud node has certain load or bandwidth limitations, the transmission of this small amount of critical information can usually be completed smoothly. This ensures that the central cloud node can obtain important event information in a timely manner for subsequent macro-level decision-making and scheduling.

[0038] The adaptive routing method for large-scale intelligent agents proposed in this application aims to solve the task processing problem of edge agents in distributed artificial intelligence systems when faced with the dual dilemma of insufficient computing resources and blocked uplink communication links.

[0039] Specifically, when an edge agent receives a task to be processed, it first performs a self-assessment. This assessment process includes two aspects: First, it determines whether it has sufficient capability to handle the task. For example, by analyzing the task's computational complexity, data volume, and its current CPU, memory, GPU, and other resource usage, it predicts whether it can complete the task within the specified time. Second, it checks the transmission status of all communication links with the central cloud node to determine if there are any congestions or faults that would prevent the task data from being efficiently uploaded to the central cloud node.

[0040] For example, in an intelligent traffic monitoring scenario, an edge agent at an intersection suddenly detects a serious traffic accident, requiring high-precision video analysis. The edge agent immediately assesses its own computing power and finds that its local processor is already running at full capacity, unable to complete the complex video analysis task within the urgent timeframe. Simultaneously, it detects that all communication links connecting to the central cloud node, including the primary fiber optic and backup wireless links, are blocked due to regional network congestion, preventing the transmission of large amounts of high-definition video data.

[0041] Faced with simultaneous bottlenecks in both computation and communication, the edge agent will not attempt to forcibly upload the task to the central cloud node, nor will it abandon the task due to insufficient capabilities. Instead, it will initiate an adaptive routing mechanism, sending resource requests to its neighboring edge agents (e.g., neighboring node 3, neighboring node 4). Upon receiving the request, these neighboring nodes will report their current resource status; for example, neighboring node 3 reports 50% CPU idle time, while neighboring node 4 reports 80% CPU idle time. Based on these reports, the edge agent will select neighboring node 4, which has the most abundant resources, as the task processing node.

[0042] Subsequently, a direct, point-to-point communication channel is established between the edge agent and the selected neighboring node 4. Through this channel, the edge agent transmits the raw high-definition video data of the accident scene to the neighboring node 4. The neighboring node 4 utilizes its ample computing resources to perform in-depth processing on the received raw video data, such as identifying the type of vehicle involved in the accident, license plate number, extent of damage, and whether anyone is trapped. After processing, the neighboring node 4 does not send back the entire processed video or a large analysis report, but instead extracts the most critical information, such as: "Time: XX o'clock XX minute, Location: XX intersection, Event: Three-vehicle rear-end collision, Preliminary judgment: People are trapped and require emergency rescue." The amount of this critical information is much smaller than the amount of raw video data.

[0043] Finally, neighboring node 4 transmits this refined critical information to the central cloud node through its relatively unobstructed communication channel with the central cloud node. Due to the significantly reduced data volume, even with some load on the communication channel between neighboring node 4 and the central cloud node, this critical information can still be delivered to the central cloud node quickly and reliably. Upon receiving the critical information, the central cloud node can immediately activate its emergency response mechanism, dispatch rescue forces, and adjust the overall traffic situation in conjunction with data from other regions.

[0044] Through the above process, the adaptive routing method of this application enables edge agents to intelligently route high-priority tasks to suitable neighboring nodes for processing when faced with complex situations such as limited resources and uplink congestion. It also significantly reduces the amount of data transmission by extracting key information, thereby ensuring that key information can be transmitted to the central cloud node in a timely and effective manner, which greatly improves the response capability and robustness of the distributed artificial intelligence system in emergency situations.

[0045] Compared to existing technologies, the advantages of this application are that, under the dual bottlenecks of edge computing and communication, it can provide an effective task processing and data transmission path, avoid system paralysis, and has higher robustness. By routing to neighboring nodes laterally for processing and transmitting only critical information, it significantly reduces end-to-end latency in task processing and information transmission, makes full use of the distributed computing resources in the edge network, improves the resource utilization efficiency of the entire system, and can adaptively adjust the task routing strategy according to the real-time changes in computing power and network status, rather than relying on preset static rules.

[0046] This application further proposes that the selected neighboring nodes process the received raw data and extract key information from the processing results, wherein the amount of key information is less than the amount of raw data. The steps also include:

[0047] When a selected neighboring node receives an information query instruction from the central cloud node, the selected neighboring node performs a preliminary identification of the information query instruction and determines the type of the information query instruction.

[0048] When the type of the information query instruction is a non-standard query, the selected neighboring node performs intent parsing on the information query instruction and extracts the query intent in the information query instruction;

[0049] The selected neighboring nodes associate the extracted query intent with the currently processed accident scene information, supplement the ambiguous information in the information query instruction, and obtain a clear information query instruction intent.

[0050] The selected neighboring nodes map the explicit information query instruction intent into a video analysis subtask and assess whether their own processing capabilities are sufficient to complete the video analysis subtask.

[0051] When the selected neighboring node assesses that its processing capabilities are insufficient to complete the video analysis subtask, the selected neighboring node generates a capability boundary report and sends it back to the central cloud node.

[0052] Specifically, when a selected neighboring node receives an information query instruction from the central cloud node, it first performs a preliminary identification of the instruction to determine its type. Information query instructions can be categorized into standardized and non-standardized queries. Standardized queries typically refer to predefined, fixed-format queries, such as querying the number of vehicles within a specific time period; non-standardized queries are more flexible and complex, potentially including natural language descriptions or fuzzy query conditions.

[0053] Specifically, when an information query instruction is determined to be a non-standard query, the selected neighboring nodes will perform intent parsing on the instruction. Intent parsing aims to accurately extract the user's true query intent from unstructured query text, such as by identifying keywords, phrases, and their semantic relationships through Natural Language Processing (NLP) techniques.

[0054] This application's solution effectively addresses the inefficiencies and inaccurate responses that may arise in basic solutions when handling non-standardized queries from central cloud nodes by introducing intelligent recognition, intent parsing, and task mapping mechanisms for information query commands. Timely generation of capability boundary reports enables central cloud nodes to receive real-time capability feedback, thereby avoiding ineffective task allocation and resource waste, and ensuring efficient operation and resource optimization of the entire system in complex and dynamic query scenarios.

[0055] In some preferred embodiments, suppose a multi-vehicle rear-end collision occurs on a highway, and the edge agent has transmitted the raw video data of the accident scene to selected neighboring nodes for processing. At this time, the central cloud node sends an information query command to the selected neighboring nodes: "Please check if any vehicles were driving in the wrong direction when the accident occurred and identify their license plates."

[0056] Upon receiving the instruction, the selected neighboring nodes first perform preliminary identification, determining it to be a non-standardized query. Then, they perform intent parsing, extracting the query intents of "finding vehicles driving in the wrong direction" and "identifying license plates." Next, the selected neighboring nodes correlate these intents with the current "multi-vehicle rear-end collision on the highway" scenario information, supplementing ambiguous information, such as limiting "driving in the wrong direction" to the lane and time period where the accident occurred, thus obtaining a clear information query instruction intent.

[0057] Subsequently, the selected neighboring node maps its explicit intent to two video analysis subtasks: "vehicle wrong-way detection" and "license plate recognition," and assesses its own processing capabilities. If the selected neighboring node possesses sufficient computing resources and the corresponding AI model to complete these two subtasks, it will directly execute the analysis and extract key information. However, if the selected neighboring node's assessment finds that its computing resources are insufficient to efficiently complete these two highly complex subtasks simultaneously, or that it lacks a high-precision license plate recognition model, it will generate a capability boundary report indicating that it cannot complete the "high-precision license plate recognition" subtask or requires more computing resources, and send this report back to the central cloud node. Upon receiving the report, the central cloud node can decide to reassign the subtask to other more capable nodes or adjust the query requirements. In this way, the selected neighboring node can intelligently respond to dynamic queries and effectively manage its own resources.

[0058] This application further proposes a more refined method for evaluating processing capabilities, which corrects the evaluation results of the processing capabilities of video analysis subtasks by dynamically adjusting the execution strategy and resource allocation.

[0059] The steps of mapping explicit information query instructions to video analysis subtasks by the selected neighboring nodes and assessing whether their processing capabilities can complete the video analysis subtasks include:

[0060] The selected neighboring node continuously acquires data quality information of the horizontal transmission link and acquires real-time data of local task load;

[0061] The selected neighboring nodes adjust the execution strategy of the video analysis subtask based on the data quality information to obtain the adjusted execution strategy. The adjustment of the execution strategy of the video analysis subtask includes: when the data quality deteriorates, executing the video analysis subtask with low data quality requirements, or reducing the analysis frame rate of the video analysis subtask with high data quality requirements.

[0062] The selected neighboring nodes adjust the computing resources allocated to the video analysis subtask based on the real-time data of the local task load, and obtain the computing resource allocation result. The adjustment of the computing resources allocated to the video analysis subtask includes: reducing the non-critical computing resources allocated to the video analysis subtask when the local task load increases, and ensuring the allocation of critical computing resources for the video analysis subtask.

[0063] The selected neighboring nodes, based on the adjusted execution strategy and computing resource allocation results, correct the evaluation results of the processing capabilities of the video analysis subtask.

[0064] Specifically, when evaluating its own processing capabilities, the selected neighboring node continuously acquires data quality information from the horizontal transmission link and real-time data on its local task load. The data quality information from the horizontal transmission link can be understood as metrics such as the clarity, integrity, and transmission latency of the video stream; these metrics directly affect the execution performance and resource requirements of the video analysis subtask. Real-time data on the local task load refers to the number of other tasks currently being processed by the selected neighboring node, CPU utilization, memory usage, etc., reflecting the node's real-time computing resource pressure.

[0065] Furthermore, the selected neighboring nodes dynamically adjust the execution strategy of the video analysis subtasks based on the acquired data quality information. For example, when data quality deteriorates, to ensure task completion, video analysis subtasks with lower data quality requirements can be executed, such as switching from high-precision recognition to low-precision recognition, or the analysis frame rate of video analysis subtasks with higher data quality requirements can be reduced, such as reducing the number of frames analyzed per second from 30 to 15, to reduce computational load. The aim is to effectively utilize existing data for analysis even when data quality is poor, avoiding task interruption or unusable results due to data quality issues.

[0066] Simultaneously, the selected neighboring nodes will adjust the computing resources allocated to the video analysis subtasks based on real-time data of the local task load. Specifically, when the local task load increases, to prioritize the execution of core tasks, the allocation of non-critical computing resources to the video analysis subtasks can be reduced, such as decreasing the number of CPU cores or memory used for auxiliary analysis or log recording. However, the allocation of critical computing resources for the video analysis subtasks will be ensured, such as ensuring that GPU resources or high-priority CPU cores used for core algorithm processing are not overwhelmed. The aim is to ensure the smooth execution of critical tasks through intelligent scheduling when resources are scarce, while optimizing overall resource utilization efficiency.

[0067] This application further proposes the following steps for the selected neighboring nodes to map explicit information query intents to video analysis subtasks:

[0068] The selected neighboring nodes obtain the data quality information of the original data;

[0069] The selected neighboring nodes obtain information about the availability of their own computing resources;

[0070] The selected neighboring nodes adjust the granularity of the video analysis subtask based on the data quality information and the availability of their own computing resources.

[0071] The selected neighboring nodes adjust the complexity of the video analysis subtask based on the data quality information and the availability of their own computing resources.

[0072] The selected neighboring nodes prioritize the video analysis subtasks based on the data quality information and their own computing resource availability information.

[0073] Specifically, when a selected neighboring node maps a clear information query instruction to a video analysis subtask, it first obtains the data quality information of the raw data. This data quality information may include, but is not limited to, video resolution, frame rate, encoding quality, noise level, and occlusion status, reflecting the availability and analysis difficulty of the raw data. Simultaneously, the selected neighboring node also obtains information on the availability of its own computing resources, such as CPU utilization, GPU load, memory usage, storage space, and network bandwidth, to understand the current node's processing capabilities.

[0074] Furthermore, the selected neighboring nodes dynamically adjust the granularity of the video analysis subtasks based on the acquired data quality information and the availability of their own computing resources. For example, when the original data quality is high and computing resources are sufficient, the granularity of the video analysis subtasks can be set to be finer, allowing for more refined target recognition or behavior analysis; while when the data quality is low or computing resources are scarce, the granularity of the subtasks can be appropriately reduced, for example, by changing from frame-by-frame analysis to keyframe analysis, to ensure the completion of basic analysis.

[0075] Furthermore, the selected neighboring nodes will adjust the complexity of the video analysis subtasks based on the data quality information and the availability of their own computing resources. For example, when dealing with high-resolution, low-noise raw data and having ample computing resources, more complex deep learning models can be executed for multi-object tracking or abnormal behavior prediction; while when the data quality is poor or resources are limited, traditional image processing algorithms or simplified models with lower computational requirements may be chosen to ensure real-time performance.

[0076] Finally, the selected neighboring nodes prioritize the video analysis subtasks based on the data quality information and their own computing resource availability. For example, subtasks with high data quality, sufficient resources, and high real-time requirements can be assigned higher priority and scheduled for execution first; while subtasks with poor data quality, scarce resources, or low importance can have their priority reduced, be processed later, or be simplified.

[0077] In some preferred embodiments, it is assumed that an edge agent is deployed in an urban traffic monitoring scenario, responsible for processing video streams from multiple cameras. When the central cloud node issues an information query command, such as "identify all vehicle collisions that occurred at a specific intersection in the past five minutes and track the driving trajectories of the vehicles involved," the selected neighboring nodes first perform intent parsing on the command to obtain the clear query intent.

[0078] Subsequently, the selected neighboring nodes will obtain data quality information about the raw video data transmitted by the intersection's cameras. For example, if the video stream suffers from reduced resolution or unstable frame rate due to network fluctuations, or if the image is blurry due to weather factors such as rain or fog, the data quality information will reflect these issues. Simultaneously, the selected neighboring nodes will also obtain information about the availability of their own computing resources, such as high GPU load or near-saturation CPU utilization.

[0079] Based on this real-time information, the selected neighboring nodes will dynamically adjust the granularity, complexity, and priority of the video analysis subtasks. Specifically:

[0080] If the original video data is of poor quality (e.g., low resolution, blurry), and the selected neighboring nodes have limited computing resources, the granularity of the "vehicle recognition" subtask may be adjusted from fine-grained vehicle model identification to coarse vehicle presence detection, and the complexity of the "trajectory tracking" subtask may be reduced, for example, from a high-precision multi-object tracking algorithm to a simplified region-based tracking algorithm. Meanwhile, the core task of "collision accident recognition" may be given higher priority to ensure its completion even under resource constraints.

[0081] Conversely, if the original video data is of extremely high quality (e.g., 4K high definition, stable frame rate) and has sufficient computing resources, the selected neighboring nodes can finer-grained the "vehicle recognition" subtask, such as recognizing the vehicle's brand, color, or even license plate, and use more complex deep learning models for high-precision multi-target tracking and behavior analysis to provide more detailed accident reports.

[0082] In this way, the selected neighboring nodes can flexibly generate and schedule video analysis subtasks according to the actual operating environment and task requirements, ensuring that information query instructions can be completed efficiently and accurately under different conditions.

[0083] This application further proposes a step whereby the selected neighboring nodes map explicit information query instructions to video analysis sub-tasks and assess whether their processing capabilities are sufficient to complete the video analysis sub-tasks, including:

[0084] The priority factor included in the selected neighbor node information query instruction;

[0085] The selected neighboring nodes obtain the data quality information of the original data;

[0086] The selected neighboring nodes obtain information about the availability of their own computing resources;

[0087] The selected neighboring nodes calculate the comprehensive priority score for each video analysis subtask based on the priority factors, the data quality information, and the availability information of their own computing resources.

[0088] The selected neighboring nodes rank the video analysis subtasks according to the overall priority score.

[0089] Specifically, the selected neighboring node information query instruction includes priority factors. These priority factors are attributes parsed from the information query instruction that indicate the importance or urgency of the task. For example, these factors may include the priority explicitly specified by the user during the query (e.g., "urgent," "high," "medium," "low"), the business type to which the task belongs (e.g., "security alert," "traffic flow analysis," "equipment fault diagnosis"), or the degree of urgency associated with a specific event (e.g., "sudden accident"). These priority factors provide a direct and crucial basis for the initial assessment of the task's importance.

[0090] The selected neighboring nodes acquire data quality information of the raw data. This data quality information can be understood as an assessment of the completeness, clarity, real-time performance, and accuracy of the raw data required for the video analysis subtask. For example, data quality information may include the video stream's resolution, frame rate, encoding format, presence of dropped or corrupted frames, and data transmission latency. High-quality data typically allows for more refined and complex analysis and may improve the reliability of the analysis results, while low-quality data may limit the depth or accuracy of the analysis.

[0091] In practical applications, the selected neighboring nodes acquire information about the availability of their own computing resources. Specifically, this information refers to the real-time status of the computing resources (such as CPU, GPU, memory, storage, network bandwidth, etc.) currently available to the selected neighboring nodes for executing video analysis subtasks.

[0092] The selected neighboring nodes calculate a comprehensive priority score for each video analytics subtask based on the priority factors, data quality information, and their own computing resource availability. This comprehensive priority score is a quantitative indicator used to comprehensively measure the relative importance and executability of each video analytics subtask. For example, a unified priority score can be obtained by assigning weights to different priority factors, data quality levels, and resource availability status, and then summing these weighted values ​​or by calculating using a multi-criteria decision model. This comprehensive evaluation method ensures the comprehensiveness and objectivity of task priority determination.

[0093] Therefore, the selected neighboring nodes sort the video analysis subtasks according to the comprehensive priority score. This means that all pending video analysis subtasks will be arranged from high to low according to their calculated comprehensive priority scores, forming an ordered task queue. Tasks with higher priority scores will be scheduled and executed first, ensuring that critical tasks can be processed in a timely manner, thereby optimizing the task execution order.

[0094] Through the above technical solution, this application enables more refined and intelligent priority management of video analytics subtasks. Compared to sorting solely based on data quality and resource availability, this application incorporates priority factors into information query commands, allowing task scheduling decisions to better align with actual business needs and urgency. This not only improves the response speed and completion efficiency of critical tasks but also optimizes the allocation of limited computing resources, ensuring that resources are prioritized for the most valuable video analytics tasks, thereby significantly enhancing the robustness and intelligence of the entire adaptive routing method for large-scale intelligent agents.

[0095] In some preferred embodiments, it is assumed that a selected neighboring node receives multiple information query instructions and maps them to multiple video analysis subtasks, for example:

[0096] 1. Subtask A: Urgently search for all abnormal behaviors within a specific area (the information query instruction includes the "urgent" priority factor).

[0097] 2. Subtask B: Routinely count traffic flow at a certain intersection (the information query command includes the "routine" priority factor).

[0098] 3. Subtask C: Analyze historical video data to optimize traffic lights (information query instructions include a "low" priority factor).

[0099] Meanwhile, the selected neighboring nodes obtain the following information:

[0100] Data quality information for the raw data: The real-time video streams of subtasks A and B are of high quality, while the historical video data of subtask C is of medium quality.

[0101] Information on the availability of computing resources: CPU and GPU resources are currently sufficient, but network bandwidth is slightly strained.

[0102] Based on the above information, the selected neighboring nodes will calculate a comprehensive priority score for each subtask. For example, a scoring model can be set up where the "urgent" priority factor has the highest weight, and high-quality data and sufficient resources will also increase the score.

[0103] Subtask A: Based on the "urgent" priority factor, high-quality data, and sufficient resources, calculate a high overall priority score, such as 70 points.

[0104] Subtask B: Based on “normal” priority factors, high-quality data, and sufficient resources, calculate a moderate overall priority score, such as 40 points.

[0105] Subtask C: Based on “low” priority factors, medium-quality data, and sufficient resources, calculate a lower overall priority score, such as 25 points.

[0106] Based on these comprehensive priority scores, the selected neighboring nodes will sort the video analysis subtasks, resulting in the following ranking: Subtask A > Subtask B > Subtask C. Therefore, Subtask A will be scheduled and executed first, ensuring the fastest possible response and handling of emergencies, while other tasks will be performed sequentially according to their importance, thus achieving more efficient and intelligent task scheduling.

[0107] This application further proposes that the step of sorting the video analysis subtasks by the selected neighboring nodes based on the comprehensive priority score includes:

[0108] The selected neighboring nodes obtain the specific computing resource requirements of video analysis subtasks with the same comprehensive priority score for different computing resources, and thus obtain resource requirement information.

[0109] The selected neighboring nodes obtain the real-time availability of their various computing resources.

[0110] The selected neighboring nodes identify a set of subtasks with resource conflicts based on the resource requirement information of video analysis subtasks with the same comprehensive priority score and the real-time availability of resources.

[0111] The selected neighboring nodes perform a refined evaluation of each subtask in the set of subtasks with resource conflicts, based on their contribution to the extraction of key information, and obtain a refined evaluation result.

[0112] The selected neighboring nodes perform a secondary sorting of the set of subtasks with resource conflicts based on the refined evaluation results, and prioritize scheduling subtasks whose contribution to the extraction of key information is greater than the prediction threshold.

[0113] Specifically, when multiple video analytics subtasks are evaluated to have the same overall priority score, the selected neighboring nodes first obtain the specific requirements of these subtasks for different types of computing resources (such as CPU, GPU, memory, storage, etc.), and simultaneously obtain their own real-time availability of various computing resources. The resource requirement information can include parameters such as the number of computing cores, memory size, storage space, and network bandwidth required for each subtask. Real-time availability reflects the actual resources that the current node can allocate. Based on this information, the selected neighboring nodes can identify which subtasks are competing for resources, meaning their total resource requirements exceed the current node's real-time availability, thus forming one or more sets of subtasks with resource conflicts.

[0114] Furthermore, for each subtask in the identified set of resource-conflicting subtasks, the selected neighboring nodes undergo a refined evaluation based on their contribution to key information extraction. The contribution to key information extraction refers to the importance of the subtask's execution in obtaining core information valuable to the central cloud node from the raw data. For example, in a video analytics scenario, the subtask of identifying a specific target person may contribute more to key information extraction than the subtask of counting background vehicles. This contribution can be quantified using pre-defined business rules, machine learning models, or expert knowledge. Through this refined evaluation, a detailed assessment result for each conflicting subtask can be obtained.

[0115] Therefore, the selected neighboring nodes perform a secondary sorting of the set of subtasks with resource conflicts based on the refined evaluation results. During this secondary sorting, subtasks whose contribution to key information extraction is greater than a preset prediction threshold are prioritized for scheduling. This prediction threshold can be dynamically adjusted according to business needs and system performance requirements to ensure that the most critical subtasks are processed first.

[0116] In some of the embodiments described above in this application, a refined evaluation is proposed for each subtask in a set of subtasks with resource conflicts, based on its contribution to the extraction of key information. However, ensuring the accuracy and adaptability of this contribution evaluation during its implementation, especially when business priorities or incident scenarios change dynamically, is a problem that needs to be solved. If the evaluation mechanism is static or inflexible, it may not accurately reflect the true contribution of different subtasks to key information in a specific context, potentially leading to improper resource allocation, affecting the timely extraction of key information and the overall system's decision-making efficiency.

[0117] To address this, this application further proposes a method for refining the evaluation of each subtask in the set of resource-conflicting subtasks by the selected neighboring nodes, based on their contribution to the extraction of key information. The steps for obtaining the refined evaluation results include:

[0118] The selected neighboring node obtains a set of business priority rules issued by the central cloud node. The set of rules includes preset contribution weights of different types of video analysis subtasks to the extraction of key information, as well as adjustment coefficients under specific accident scenarios.

[0119] The selected neighboring node modifies the adjustment coefficients in the rule set according to the business priority rule set and in combination with the current accident scenario type, to obtain the modified business priority rule set.

[0120] The selected neighboring nodes calculate the contribution score of each subtask in the set of subtasks with resource conflicts based on the modified business priority rule set.

[0121] The selected neighboring nodes will extract the contribution score of the key information and compare it with the historical evaluation data issued by the central cloud node to identify the deviation of the evaluation results;

[0122] When the deviation of the evaluation result is found to exceed a preset threshold, the selected neighboring node sends an evaluation deviation report to the central cloud node and receives an evaluation correction instruction issued by the central cloud node.

[0123] Specifically, the set of business priority rules can be understood as a series of strategies uniformly managed and distributed by the central cloud node. These strategies aim to guide edge agents in evaluating the importance of different sub-tasks for extracting key information when processing video analysis sub-tasks. This set of rules typically includes preset contribution weights for different types of video analysis sub-tasks. For example, the weight of the "vehicle recognition" sub-task may be higher than that of the "background noise analysis" sub-task, because vehicle recognition is usually directly related to key information in accident scenarios. Furthermore, the set of rules also includes adjustment coefficients for specific accident scenarios. For example, in a "traffic accident" scenario, the weight of the "license plate recognition" sub-task may be significantly increased, while in a "crowd gathering" scenario, the weight of the "face recognition" sub-task may be increased. These adjustment coefficients allow the evaluation process to be dynamically adjusted according to actual business needs and scenario changes.

[0124] The selected neighboring node adjusts the adjustment coefficients in the business priority rule set based on the current accident scenario type, resulting in a revised business priority rule set. Specifically, when a selected neighboring node handles a specific accident scenario (e.g., a fire scene, traffic congestion, or suspicious person loitering), it identifies adjustment coefficients related to that scenario from the business priority rule set. For example, if the current scenario is identified as a "fire scene," the adjustment coefficients related to sub-tasks such as "smoke detection" and "flame recognition" are increased to reflect the higher priority and critical information contribution of these sub-tasks in that scenario. This ensures that the evaluation of sub-task contribution is real-time and highly relevant to the current business scenario.

[0125] In practical applications, the selected neighboring nodes calculate the key information extraction contribution score for each subtask in the set of subtasks with resource conflicts, based on the revised business priority rule set. This contribution score is a quantitative indicator used to measure the potential contribution of each subtask to extracting key information of the accident scenario under the current revised business priority rules. The calculation method may involve multiplying a preset contribution weight by a revised adjustment coefficient, and may also combine other factors, such as the real-time processing difficulty of the subtask and the amount of data, to obtain a comprehensive contribution score.

[0126] Furthermore, the selected neighboring nodes compare the contribution score of the extracted key information with the historical evaluation data issued by the central cloud node to identify deviations in the evaluation results. Historical evaluation data may include contribution scores from similar past tasks, the actual efficiency of key information extraction, and feedback or correction records from the central cloud node regarding these evaluations. By comparing these data, it can be determined whether the currently calculated contribution score differs significantly from historical trends or expected values. For example, if the contribution score of a subtask is far below the historical average or deviates significantly from previous evaluation results from the central cloud node, a deviation may exist.

[0127] When the deviation of the evaluation result exceeds a preset threshold, the selected neighboring node sends an evaluation deviation report to the central cloud node and receives an evaluation correction instruction from the central cloud node. The preset threshold defines the level of deviation that requires attention and intervention. Once the deviation exceeds this threshold, it indicates that the current evaluation may be inaccurate or inconsistent with the central cloud node's global strategy. In this case, the selected neighboring node generates a report detailing the deviation and sends it to the central cloud node. After receiving the report, the central cloud node can adjust the evaluation model or rules based on its global perspective and more comprehensive data, and issue an evaluation correction instruction to guide the selected neighboring node in optimizing its subsequent evaluation process.

[0128] This application's solution effectively addresses the issue of maintaining the accuracy and adaptability of key information extraction contribution assessment in dynamic environments by introducing a dynamic set of business priority rules and a historical evaluation data comparison mechanism. Specifically, firstly, by acquiring the set of business priority rules issued by the central cloud node, basic contribution weights and scene adjustment coefficients are set for different types of video analysis subtasks, providing a unified and configurable benchmark for evaluation. Secondly, selected neighboring nodes can adjust these adjustment coefficients in real time based on the type of accident scenario being processed, ensuring that the contribution assessment closely aligns with actual business needs and scenario changes, avoiding the limitations of static evaluation. Therefore, the calculated key information extraction contribution score can more accurately reflect the true value of the subtask in a specific context. In some preferred embodiments, suppose an edge agent is monitoring a large transportation hub and receives an information query instruction from the central cloud node, requesting analysis of real-time video streams in a certain area to identify potential traffic accident risks. During processing, selected neighboring nodes identify multiple video analysis subtasks, such as "vehicle speed detection," "abnormal parking identification," "pedestrian density analysis," and "license plate recognition," and these subtasks have computational resource conflicts.

[0129] To conduct a refined evaluation of these conflicting subtasks, the selected neighboring nodes first obtain a set of business priority rules from the central cloud node. This set of rules may stipulate that "vehicle speed detection" and "abnormal parking identification" have higher preset contribution weights in general traffic scenarios, while "pedestrian density analysis" and "license plate recognition" have relatively lower weights. Simultaneously, the set of rules also includes adjustment coefficients for specific scenarios such as "traffic congestion" and "traffic accidents."

[0130] The current accident scenario is identified as "traffic congestion." The selected neighboring nodes adjust the adjustment coefficients in the rule set based on this scenario type. For example, in a "traffic congestion" scenario, the adjustment coefficients for "abnormal parking identification" and "vehicle speed detection" might be further increased, as they are crucial for quickly identifying the causes of congestion and potential risks; while the adjustment coefficient for "license plate recognition" might also be moderately increased to track specific vehicles when congestion occurs.

[0131] Based on the revised set of business priority rules, the selected neighboring nodes calculate the contribution score for extracting key information for each conflicting subtask. For example, "abnormal parking identification" may receive the highest contribution score, followed by "vehicle speed detection", then "license plate recognition", and finally "pedestrian density analysis".

[0132] Subsequently, the selected neighboring nodes compare these calculated contribution scores with the historical evaluation data sent by the central cloud node. Assuming historical data shows that in past "traffic congestion" scenarios, the actual contribution score for "abnormal parking identification" is typically between 0.8 and 0.9. If the currently calculated contribution score for "abnormal parking identification" is only 0.6, and this deviation exceeds the preset threshold of 0.1, the selected neighboring nodes will identify a bias in the evaluation result.

[0133] In this scenario, the selected neighboring nodes immediately generate an evaluation bias report detailing the difference between the current evaluation score and historical data for the "abnormal parking recognition" subtask, and send it to the central cloud node. Upon receiving the report, the central cloud node may analyze the causes of the bias, such as whether the rule set needs updating or whether there are problems with the edge agent's evaluation model. The central cloud node then issues evaluation correction instructions, for example, instructing the selected neighboring nodes to update their local adjustment coefficient parameters or provide a more accurate evaluation model to ensure more accurate subsequent evaluations.

[0134] This application further proposes a step to identify deviations in the evaluation results by comparing the contribution scores extracted from the key information by the selected neighboring nodes with the historical evaluation data issued by the central cloud node.

[0135] Before the selected neighboring nodes compare the contribution score extracted from the key information with the historical evaluation data issued by the central cloud node, they perform a data integrity check on the historical evaluation data, identify and mark records with missing data.

[0136] The selected neighboring nodes perform data consistency checks on the historical evaluation data, identify and mark records with abnormal data, including records that exceed a preset value range or records that deviate significantly from the same type of data;

[0137] The selected neighboring nodes identify and mark outdated records based on the generation timestamp of the historical evaluation data. Outdated records refer to records whose generation time exceeds a preset time limit.

[0138] The selected neighboring nodes filter and clean the historical evaluation data based on the records with missing data, records with abnormal data, and records with outdated data, removing or correcting data that does not meet the requirements, to obtain cleaned historical evaluation data.

[0139] The selected neighboring nodes construct a dynamic comparison benchmark based on the cleaned historical evaluation data. The dynamic comparison benchmark can reflect the evaluation trend and fluctuation range under the current business scenario.

[0140] The selected neighboring nodes extract the contribution score of the key information and compare it with the dynamic comparison benchmark to identify the deviation of the evaluation result.

[0141] Specifically, before comparison, the selected neighboring nodes preprocess the received historical evaluation data. Data integrity checks aim to identify and flag null values, incomplete fields, or records in the historical data to ensure data integrity. Data consistency checks are used to identify logical conflicts or outliers within the data; for example, an evaluation score may exceed a reasonable range, or there may be a significant statistical deviation compared to similar historical data. Furthermore, by examining the generation timestamps of the historical evaluation data, outdated data generated beyond a preset period can be identified and flagged, as outdated data may not accurately reflect the current business situation and evaluation trends.

[0142] After completing the above checks, the selected neighboring nodes will filter and clean the historical evaluation data based on the identified records with missing data, abnormal data, and outdated data. This process includes removing data records that do not meet the requirements or correcting data that can be corrected, thereby obtaining high-quality, reliable cleaned historical evaluation data.

[0143] Furthermore, based on the cleaned historical evaluation data, a dynamic comparison benchmark is constructed using the selected neighboring nodes. This dynamic benchmark differs from a static threshold; it adaptively adjusts according to the actual trends and fluctuations of historical data, thus more accurately reflecting the evaluation standards in the current business scenario. For example, this benchmark could be an average value or statistical interval based on a sliding window, capable of capturing the changing patterns of the evaluation data in real time.

[0144] Finally, the selected neighboring nodes extract contribution scores from the currently calculated key information and compare them with the constructed dynamic comparison benchmark. In this way, significant biases in the evaluation results can be identified more accurately, avoiding misjudgments caused by historical data quality issues.

[0145] This application's solution effectively addresses the inaccuracies that may arise from directly comparing raw historical data by introducing preprocessing of historical evaluation data and the construction of a dynamic comparison benchmark. Specifically, data integrity checks, data consistency checks, and data obsolescence checks ensure that the historical data used for comparison is high-quality, reliable, and timely. Through screening and cleaning, potential noise and erroneous information are eliminated, making subsequent evaluations more accurate. Based on this, a dynamic comparison benchmark is constructed, so that the evaluation no longer relies on fixed, potentially unsuitable, static thresholds, but can adaptively adjust according to actual historical trends and fluctuation ranges. Therefore, when the contribution score of key information extraction is compared with this dynamic comparison benchmark, it can more accurately reflect the true evaluation deviation, thus providing a more reliable decision-making basis for the central cloud node.

[0146] In some preferred embodiments, it is assumed that the selected neighboring node is processing a video analysis subtask in a traffic monitoring scenario and calculating the current key information extraction contribution score. The central cloud node distributes historical evaluation data from the past week for comparison. If these raw historical data are used directly for comparison, the following problems may be encountered: the historical data records for some days are incomplete, such as missing evaluation scores for specific time periods; on a certain day, due to system failure, an evaluation score far exceeding the normal range is recorded, or there is a significant inconsistency with the data before and after; the historical data includes records from three months ago, while traffic conditions and model performance have changed significantly, and this outdated data is no longer of reference value.

[0147] According to the scheme in this application, the selected neighboring nodes will first process these historical evaluation data before comparison. First, a data integrity check will be performed to identify and mark records with missing data.

[0148] Subsequently, the selected neighboring nodes filter and clean the historical assessment data based on these tags. This results in high-quality, cleaned historical assessment data that reflects the current situation.

[0149] Next, the selected neighboring nodes construct a dynamic comparison benchmark based on this cleaned historical assessment data.

[0150] Finally, the selected neighboring nodes extract the contribution score of the currently calculated key information and compare it with this dynamic comparison benchmark. If the current score exceeds the normal fluctuation range of the dynamic comparison benchmark, a deviation in the evaluation result is identified, and a decision is made on whether to send an evaluation deviation report to the central cloud node based on the degree of deviation. In this way, misjudgments caused by problems with the quality of the original historical data are avoided, ensuring the accuracy and effectiveness of deviation identification.

[0151] The step of selecting neighboring nodes comparing the contribution score of the key information extraction with the dynamic comparison benchmark to identify deviations in the evaluation results further includes: the selected neighboring nodes continuously monitoring the real-time fluctuations of the dynamic comparison benchmark and obtaining the short-term historical fluctuation range of the dynamic comparison benchmark; the selected neighboring nodes identifying whether there are occasional or short-term abnormal fluctuations in the dynamic comparison benchmark based on the real-time fluctuations and the short-term historical fluctuation range; when occasional or short-term abnormal fluctuations are identified in the dynamic comparison benchmark, the selected neighboring nodes temporarily correct the dynamic comparison benchmark to generate a stable comparison benchmark; the selected neighboring nodes comparing the contribution score of the key information extraction with the stable comparison benchmark to identify deviations in the evaluation results.

[0152] Specifically, the selected neighboring nodes continuously monitor the real-time fluctuations of the dynamic benchmark. This means that the selected neighboring nodes acquire the numerical changes of the dynamic benchmark in real time through periodic sampling or event-driven methods. For example, they can record continuous changes in the benchmark value over a recent period. Simultaneously, the selected neighboring nodes acquire the short-term historical fluctuation range of the dynamic benchmark. This can be understood as collecting and analyzing statistical data such as the maximum, minimum, average, and standard deviation of the dynamic benchmark within a relatively short time window (e.g., the past few minutes, hours, or day) to establish an interval reflecting its normal fluctuation range. The purpose is to provide a reference for subsequent identification of abnormal fluctuations.

[0153] The selected neighboring nodes identify whether there are occasional or short-term abnormal fluctuations in the dynamic comparison benchmark based on the real-time fluctuation data and the short-term historical fluctuation range. Specifically, statistical methods can be used, such as comparing the real-time fluctuation data with the short-term historical fluctuation range. If the real-time fluctuation value significantly exceeds the upper or lower limits of the historical fluctuation range within a short period, or if the rate of change far exceeds the historical average rate of change, it can be judged as an occasional or short-term abnormal fluctuation. For example, algorithms such as moving averages and exponential smoothing can be used to detect anomalies. The purpose is to distinguish between genuine trend changes and temporary noise interference.

[0154] In practical applications, when sporadic, short-term abnormal fluctuations are identified in the dynamic comparison benchmark, the selected neighboring nodes temporarily correct the benchmark to generate a stable benchmark. Temporary correction can involve using filtering algorithms (such as Kalman filtering or median filtering) or smoothing algorithms (such as moving averages or exponentially weighted moving averages) to process the current dynamic comparison benchmark, eliminating or reducing the impact of abnormal fluctuations and restoring it to a stable state closer to its true trend. For example, if a benchmark value at a certain point in time suddenly spikes or drops, but the data before and after it are within the normal range, it can be replaced with the average of the data before and after it, or corrected using interpolation. The purpose is to provide a more stable and reliable benchmark, avoiding erroneous judgments caused by instantaneous anomalies in the benchmark itself.

[0155] Therefore, the selected neighboring nodes compare the contribution scores of the extracted key information with the stabilized comparison benchmark to identify deviations in the evaluation results. This means that when performing the final deviation identification, the original dynamic comparison benchmark is no longer used directly, but a more stable benchmark that has been temporarily modified, thereby ensuring the accuracy and reliability of the comparison.

[0156] In some preferred embodiments, a specific example is given below. Assume that in a certain accident scenario, selected neighboring nodes are processing a series of video analysis sub-tasks and continuously calculating the contribution score for extracting key information for each sub-task. Simultaneously, the selected neighboring nodes construct a dynamic comparison benchmark based on historical evaluation data to identify evaluation biases in these contribution scores.

[0157] Secondly, referring to Figure 2 This application further proposes an adaptive routing system for large-scale intelligent agents, which includes:

[0158] The information acquisition and evaluation module 201 is used to acquire information about the task to be processed by a certain edge agent, and evaluate whether the edge agent's ability to process the task to be processed is greater than the predicted processing threshold, and at the same time evaluate whether the transmission status of all communication links between the edge agent and the central cloud node is blocked.

[0159] The neighbor node selection module 202 is used to send a resource request to a neighbor node and receive the resource status reported by the neighbor node when the edge agent's ability to process the task to be processed is greater than the predicted processing threshold and the transmission status of all communication links is blocked. Based on the resource status, the edge agent selects a neighbor node as the processing node for the task to be processed and determines the selected neighbor node.

[0160] The direct communication channel establishment and data transmission module 203 is used to establish a direct communication channel between the edge agent and the selected neighboring node, and to transmit the raw data of the task to be processed to the selected neighboring node;

[0161] The data processing and key information extraction module 204 is used for the selected neighboring node to process the received raw data and extract key information from the processing result, wherein the amount of key information is less than the amount of raw data.

[0162] The key information transmission module 205 is used for the selected neighboring node to transmit the key information to the central cloud node through the communication channel that exists between it and the central cloud node.

[0163] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An adaptive routing method for large-scale intelligent agents, characterized in that, include: The system acquires information about a task to be processed by an edge agent and assesses whether the edge agent's ability to process the task is greater than the predicted processing threshold. At the same time, it assesses whether the transmission status of all communication links between the edge agent and the central cloud node is blocked. When the edge agent's ability to process the task to be processed is greater than the predicted processing threshold, and the transmission status of all communication links is blocked, the edge agent sends a resource request to the neighboring node and receives the resource status reported by the neighboring node. Based on the resource status, it selects a neighboring node as the processing node for the task to be processed and determines the selected neighboring node. The edge agent establishes a direct communication channel with the selected neighboring node and transmits the raw data of the task to be processed to the selected neighboring node; The selected neighboring node processes the received raw data and extracts key information from the processing result. The amount of key information is less than the amount of raw data. The selected neighboring nodes transmit the key information to the central cloud node through the communication channel that exists between them and the central cloud node.

2. The adaptive routing method for large-scale intelligent agents according to claim 1, characterized in that, The step of the selected neighboring node processing the received raw data and extracting key information from the processing result, wherein the amount of key information is less than the amount of raw data, further includes: When a selected neighboring node receives an information query instruction from the central cloud node, the selected neighboring node performs a preliminary identification of the information query instruction and determines the type of the information query instruction. When the type of the information query instruction is a non-standard query, the selected neighboring node performs intent parsing on the information query instruction and extracts the query intent in the information query instruction; The selected neighboring nodes associate the extracted query intent with the currently processed accident scene information, supplement the ambiguous information in the information query instruction, and obtain a clear information query instruction intent. The selected neighboring nodes map the explicit information query instruction intent into a video analysis subtask and assess whether their own processing capabilities are sufficient to complete the video analysis subtask. When the selected neighboring node assesses that its processing capabilities are insufficient to complete the video analysis subtask, the selected neighboring node generates a capability boundary report and sends it back to the central cloud node.

3. The adaptive routing method for large-scale intelligent agents according to claim 2, characterized in that, The steps of mapping explicit information query instructions to video analysis subtasks by the selected neighboring nodes and assessing whether their processing capabilities can complete the video analysis subtasks include: The selected neighboring node continuously acquires data quality information of the horizontal transmission link and acquires real-time data of local task load; The selected neighboring nodes adjust the execution strategy of the video analysis subtask based on the data quality information to obtain the adjusted execution strategy. The adjustment of the execution strategy of the video analysis subtask includes: when the data quality deteriorates, executing the video analysis subtask with low data quality requirements, or reducing the analysis frame rate of the video analysis subtask with high data quality requirements. The selected neighboring nodes adjust the computing resources allocated to the video analysis subtask based on real-time data of local task load to obtain computing resource allocation results. The adjustment of computing resources allocated to the video analysis subtask includes: reducing non-critical computing resources allocated to the video analysis subtask when the local task load increases, while ensuring the allocation of critical computing resources for the video analysis subtask. The selected neighboring nodes, based on the adjusted execution strategy and computing resource allocation results, correct the evaluation results of the processing capabilities of the video analysis subtask.

4. The adaptive routing method for large-scale intelligent agents according to claim 3, characterized in that, The steps by which the selected neighboring nodes map explicit information query intents to video analysis subtasks include: The selected neighboring nodes obtain the data quality information of the original data; The selected neighboring nodes obtain information about the availability of their own computing resources; The selected neighboring nodes adjust the granularity of the video analysis subtask based on the data quality information and the availability of their own computing resources. The selected neighboring nodes adjust the complexity of the video analysis subtask based on the data quality information and the availability of their own computing resources. The selected neighboring nodes prioritize the video analysis subtasks based on the data quality information and their own computing resource availability information.

5. The adaptive routing method for large-scale intelligent agents according to claim 4, characterized in that, The step of prioritizing the video analysis subtasks based on the data quality information and the availability of their own computing resources by the selected neighboring nodes includes: The priority factor included in the selected neighbor node information query instruction; The selected neighboring nodes obtain the data quality information of the original data; The selected neighboring nodes obtain information about the availability of their own computing resources; The selected neighboring nodes calculate the comprehensive priority score for each video analysis subtask based on the priority factors, data quality information, and their own computing resource availability information. The selected neighboring nodes rank the video analysis subtasks according to the overall priority score.

6. The adaptive routing method for large-scale intelligent agents according to claim 5, characterized in that, The step of sorting the video analysis subtasks by the selected neighboring nodes based on the comprehensive priority score further includes: The selected neighboring nodes obtain the specific computing resource requirements of video analysis subtasks with the same comprehensive priority score for different computing resources, and thus obtain resource requirement information. The selected neighboring nodes obtain the real-time availability of their various computing resources. The selected neighboring nodes identify a set of subtasks with resource conflicts based on the resource requirements of video analysis subtasks with the same comprehensive priority score and the real-time availability of resources. The selected neighboring nodes then perform a refined evaluation of each subtask in the set of subtasks with resource conflicts based on its contribution to the extraction of key information, and obtain a refined evaluation result. Based on the refined evaluation results, the selected neighboring nodes perform a secondary sorting of the set of subtasks with resource conflicts, and prioritize scheduling subtasks whose contribution to the extraction of key information is greater than the prediction threshold.

7. The adaptive routing method for large-scale intelligent agents according to claim 6, characterized in that, The steps for selecting neighboring nodes to perform a refined evaluation of each subtask in the set of subtasks with resource conflicts, based on their contribution to the extraction of key information, and to obtain the refined evaluation results include: The selected neighboring node obtains a set of business priority rules issued by the central cloud node. The set of business priority rules includes preset contribution weights of different types of video analysis subtasks to the extraction of key information, as well as adjustment coefficients under specific accident scenarios. The selected neighboring node modifies the adjustment coefficients in the rule set according to the business priority rule set and in combination with the current accident scenario type, to obtain the modified business priority rule set. The selected neighboring nodes calculate the contribution score of each subtask in the set of subtasks with resource conflicts based on the modified business priority rule set. The selected neighboring nodes will extract key information contribution scores and compare them with historical evaluation data issued by the central cloud node to identify deviations in the evaluation results. When the deviation of the evaluation result is found to exceed a preset threshold, the selected neighboring node sends an evaluation deviation report to the central cloud node and receives an evaluation correction instruction issued by the central cloud node.

8. The adaptive routing method for large-scale intelligent agents according to claim 7, characterized in that, The step of selecting neighboring nodes to extract key information contribution scores and compare them with historical evaluation data issued by the central cloud node to identify deviations in the evaluation results includes: Before the selected neighboring nodes compare the contribution score extracted from the key information with the historical evaluation data issued by the central cloud node, they perform a data integrity check on the historical evaluation data, identify and mark records with missing data. The selected neighboring nodes perform data consistency checks on historical evaluation data, identify and mark records with abnormal data, including records that exceed a preset value range or records that deviate significantly from similar data. The selected neighboring nodes identify and mark outdated records based on the generation timestamp of the historical evaluation data. Outdated records refer to records whose generation time exceeds a preset time limit. The selected neighboring nodes filter and clean the historical evaluation data based on the records with missing data, records with abnormal data, and records with outdated data, removing or correcting data that does not meet the requirements, to obtain cleaned historical evaluation data. The selected neighboring nodes construct a dynamic comparison benchmark based on the cleaned historical evaluation data. The dynamic comparison benchmark can reflect the evaluation trend and fluctuation range under the current business scenario. The selected neighboring nodes extract key information contribution scores and compare them with dynamic comparison benchmarks to identify deviations in the evaluation results.

9. The adaptive routing method for large-scale intelligent agents according to claim 8, characterized in that, The step of selecting neighboring nodes to extract contribution scores from key information and compare them with dynamic comparison benchmarks to identify deviations in the evaluation results further includes: The selected neighboring nodes continuously monitor the real-time fluctuations of the dynamic comparison benchmark and obtain the short-term historical fluctuation range of the dynamic comparison benchmark. The selected neighboring nodes identify whether there are occasional or short-term abnormal fluctuations in the dynamic comparison benchmark based on the real-time fluctuation situation and the short-term historical fluctuation range. When an occasional or short-term abnormal fluctuation is identified in the dynamic comparison benchmark, the selected neighboring node temporarily corrects the dynamic comparison benchmark to generate a stable comparison benchmark. The selected neighboring nodes compare the contribution score of the key information extraction with the smoothed comparison benchmark to identify the deviation of the evaluation results.

10. An adaptive routing system for large-scale intelligent agents, characterized in that, The system includes: The information acquisition and evaluation module is used to acquire information about the task to be processed by a certain edge agent, and evaluate whether the edge agent's ability to process the task is greater than the predicted processing threshold. At the same time, it evaluates whether the transmission status of all communication links between the edge agent and the central cloud node is blocked. The neighbor node selection module is used to send a resource request to a neighbor node and receive the resource status reported by the neighbor node when the edge agent's ability to process the task to be processed is greater than the predicted processing threshold and the transmission status of all communication links is blocked. Based on the resource status, the edge agent selects a neighbor node as the processing node for the task to be processed and determines the selected neighbor node. The direct communication channel establishment and data transmission module is used to establish a direct communication channel between the edge agent and the selected neighboring node, and to transmit the raw data of the task to be processed to the selected neighboring node; The data processing and key information extraction module is used by the selected neighboring node to process the received raw data and extract key information from the processing result, wherein the amount of key information is less than the amount of raw data. The critical information transmission module is used by the selected neighboring node to transmit the critical information to the central cloud node through the communication channel that exists between the neighboring node and the central cloud node.

Images